Kan AI diepe nepvideo's met hogere nauwkeurigheid dan menselijke experts in realtime detecteren ?

AI-systemen analyseren nu micro-expressies, lichtinconsistenties en biologische signalen om AI-gegenereerde gezichten te identificeren. Platforms zoals Microsoft Video Authenticator kunnen synthetische inhoud markeren voordat deze zich verspreidt. Deze wapenwedloop met generatieve videotechnologie is cruciaal voor het bestrijden van desinformatie. Nauwkeurigheidspercentages overtreffen in gecontroleerde studies getrainde onderzoekers. Real-time API-gebaseerde detectie is al geïmplementeerd.

Background

Current AI systems analyze micro-expressions, lighting inconsistencies, biological signals, and subtle artifacts in facial expressions or blinking patterns to flag synthetic content. State-of-the-art models—including EfficientNet, Vision Transformers, and specialized deepfake detectors (e.g., DFDC winners)—often exceed untrained human observers in controlled tests. Platforms such as Microsoft Video Authenticator demonstrate real-time API-based detection already in limited deployments. Benchmarks like the Deepfake Detection Challenge (DFDC) report higher accuracy compared to human experts on curated datasets; however, performance drops in unconstrained, real-world conditions due to factors such as latency constraints, adversarial attacks, and generalization gaps across unseen generation methods (e.g., diffusion models). The ongoing arms race with generative video technology underscores the need for continued advances in both detection and generation robustness.